A CRM information intelligent processing system and method based on natural language dialogue

Abstract

The application relates to the technical field of natural language processing, and discloses a CRM information intelligent processing method and system based on natural language dialogue; the application accesses multiple mainstream chat software and establishes unified message standardization and identity permission mapping, so that employees can complete customer information input, follow-up and query in a natural language in a commonly used chat environment, do not need to frequently switch systems, and can gradually complete business fields in line with natural communication habits in combination with multi-modal content analysis and multi-round conversation incremental semantic recognition; in addition, the three-stage verification mechanism of draft preview, manual confirmation and formal writing is matched, the influence of recognition deviation on data is significantly reduced, multi-dimensional statistical and penetrating inquiry of business data are supported, and local deployment, end-side sensitive information desensitization and full-link traceability are supplemented; on the basis of improving operation efficiency and data quality, data safety compliance and traceability are comprehensively strengthened.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically discloses a CRM information intelligent processing system and method based on natural language dialogue. Background Technology

[0002] As enterprises deepen their digital transformation, customer relationship management (CRM) systems have become core management tools for enterprises to conduct business activities around the customer lifecycle, accumulate customer data, and keep follow-up records. However, traditional CRM systems usually require employees to frequently log in to independent business pages and enter and update customer information in a structured way by filling out forms and manually selecting fields. This process is not only cumbersome and disconnected from the real-time communication scenarios that employees rely on in their daily work, but it is also prone to customer data distortion and missing follow-up records due to information transmission delays or secondary entry errors.

[0003] In the prior art, document CN111353818B discloses a "Segmented Processing Method for Multiple Membership Enrollments Based on CRM," which includes a CRM system. It uses a PC-based binding application unit, a withdrawal management application unit, a re-entry management application unit, and a data processing application unit in conjunction with the CRM system as the basis for segmented processing of multiple membership enrollments. In the binding application unit, when a user initially joins on the corresponding e-commerce platform, they bind their mobile phone number. After this mobile phone number corresponds to a segment of membership information, the CRM system generates membership data and benefits. In the withdrawal management application unit, when a user withdraws, they unbind their mobile phone number, and the CRM system archives the membership data and benefits information before unbinding. In the re-entry management application unit, the CRM system continues the membership data and benefits information previously archived in the withdrawal management application unit.

[0004] Document CN115473939B discloses a “Page Communication Method and Apparatus Based on Customer Relationship Management (CRM) System”. The method includes acquiring communication data to be processed and encapsulating the communication data to be processed into target communication data in a preset data format. The preset data format includes a first application layer data format negotiated by multiple embedded pages, where each embedded page represents a user page embedded in the CRM system homepage. The target communication data is then sent to the CRM system homepage, causing the CRM system homepage to broadcast and forward the target communication data to multiple embedded pages. The embedded pages then compare the target communication data with the first application layer data format and a stored second application layer data format to obtain user page information from the target communication data.

[0005] In addition, the patent with publication number CN112955894A involves semantic inference in customer relationship management, the core of which is to parse CRM event text into descriptive logic triples and write them into the enterprise knowledge graph; the patent US11210300B2 involves training a model using historical unstructured data and mapping it to structured fields using specific n-grams and weights; the patent US9467542B1 mainly involves mobile phone contact management, automatic contact creation and call routing.

[0006] While the aforementioned existing technologies each have their own characteristics, most only support a single instant messaging platform. The message formats and account systems of different platforms are independent, making it impossible to achieve unified mapping and access control across platforms. Furthermore, semantic recognition often adopts a single-turn instruction-based processing, requiring employees to fully express all fields at once. If a field is missing, an error is directly reported instead of being guided to complete it. It also cannot utilize multi-turn conversation context for incremental recognition. More importantly, the AI ​​recognition results of existing solutions have direct or indirect permissions to write to the database, lacking intermediate verification and confirmation steps, which can easily lead to data contamination due to recognition errors. At the same time, they lack the ability to locate fine-grained evidence for multimodal content (images, voice, files), making it impossible to trace the specific source of CRM fields. Business data analysis relies on manual statistical summarization, resulting in delayed reports that cannot penetrate from indicators to the original communication context. Cloud deployment poses a risk of data leakage and lacks the ability to trace the entire dialogue chain and pre-anonymize sensitive information. Summary of the Invention

[0007] This invention provides a CRM information intelligent processing system and method based on natural language dialogue, which can solve the problems mentioned in the background art.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution, more specifically, a CRM information intelligent processing method based on natural language dialogue, comprising: S1. Connect to various mainstream chat software through the adapter interface, receive various messages sent by employees, perform standardized message format conversion, bind and map chat accounts with CRM employee accounts, and filter duplicate messages; S2. Perform multimodal text parsing on the content of different message types, converting images, voice, and files into parsing evidence with original local references, time segments, page numbers, or image area locations, and perform comprehensive identification based on the parsing evidence, current session context, employee identity, and CRM existing data. Extract CRM business fields and identify employee operation intentions. When the identification results have missing fields, conflicting evidence, or confidence levels below the threshold, generate natural language supplementary questions and send them through the original session to guide employees to supplement information. S3. Generate an unconfirmed action plan based on the extracted business fields. The action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependency and idempotent key. Retrieve matching candidate customers in the CRM database. After completing customer deduplication verification, generate a structured CRM operation draft of the corresponding operation type. The unconfirmed action plan is generated and output by the semantic processing model. The semantic processing model has no right to modify the draft status or call the CRM database write interface. S4. Feedback the operation draft to the original chat session, receive modification, confirmation, or cancellation instructions sent by employees via natural language, update the draft content, and the confirmation instruction must simultaneously meet the following conditions: from the same enabled robot connection, from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies, and atomically executes the confirmation status change, formal data writing, audit record generation, and structured receipt generation in the same business transaction. If any step fails, the entire process is rolled back, and no partially written formal objects are generated. After writing, a structured receipt based on the actual execution result is generated, which includes the specific objects created or updated and the follow-up time. The structured receipt is generated by the CRM Core based on the actual writing result. The semantic processing model has no right to generate or modify the receipt content, nor does it have the right to generate a conclusion indicating that the writing was successful. S5 aggregates CRM business data by company, department, and employee, generates periodic business analysis reports, and stores data of all operation nodes to build a data traceability chain. For duplicate delivery of original messages, it uses a combination of platform identifier and message unique identifier to deduplicate, and only reuses existing drafts. For duplicate sending of confirmation messages, it uses a combination of local instance, robot connection, and confirmation message unique identifier to deduplicate, and only returns the first execution result, without writing it repeatedly.

[0009] Furthermore, in step S1, the standardized and transformed message includes platform identifier, unique message identifier, session identifier, sender account, sending time, message type, and content attachment information; after account binding and mapping, the corresponding employee's CRM data operation permissions are loaded synchronously to limit the range of customer data that employees can access; the combination of platform identifier and unique message identifier is used as the deduplication key to intercept duplicate messages.

[0010] Furthermore, in step S2, messages of image, voice, and file types are converted into text content through text recognition, speech-to-text transcription, and document parsing, respectively, while retaining the original local references, timestamps or regional coordinates of the recognized segments, and page number information. These are then concatenated with the plain text content in the conversation in chronological order. The semantically extracted business fields include customer name, contact information, contact method, customer stage, needs, budget, purchasing intention, follow-up content, and follow-up time. The recognized operation intentions include creating a customer, supplementing customer information, adding a contact, recording business events, arranging follow-ups, and querying customer information.

[0011] Furthermore, in step S3, customer matching adopts a two-level processing mode of recall and fine sorting. First, candidate customers are recalled based on key fields, and then the matching score is calculated by weighting multiple features. According to the matching score, automatic association, in-session manual confirmation, or new customer processing is performed in stages. The structured operation draft contains the customer profile to be written, contact information, follow-up records, and business event details, and the field format legality is verified. The action dependency relationship in the unconfirmed action plan supports the arrangement of multiple actions into the same transaction batch, such as creating a customer first, and then creating a contact and follow-up action under that customer.

[0012] Furthermore, in step S4, the operation draft is pushed to the original chat session in the form of structured cards or natural language summaries; when an employee's modification instruction is received, the corresponding fields and values ​​are parsed and modified, the draft is updated, and then pushed back for preview; if no confirmation instruction is received, the draft is only temporarily cached and not written to the CRM database; after writing is completed, the execution result is returned to the original chat session; the semantic processing model does not hold a CRM database write certificate; the CRM Core runs independently of the semantic processing model, and receives action plans and performs verification and submission through the internal service interface as a controlled service.

[0013] Furthermore, in step S5, the statistical business data includes the number of new customers, changes in customer stages, follow-up completion status, overdue matters, and dynamics of high-level customers; it supports drill-down from statistical indicators to the corresponding customers, business records, and original chat messages; the node data stored throughout the process includes original messages, multimodal parsing results and evidence location, semantic recognition process, customer matching details, draft versions, confirmation records, and final write results.

[0014] Furthermore, the method also supports a customer-level-driven periodic follow-up mechanism: a unique pending periodic action is maintained for each customer who has enabled periodic follow-up, and the customer level corresponds to a configurable follow-up frequency; the system recalculates the next countdown based on the actual completion time only after the employee has clearly completed the follow-up and submitted new communication content or manual instructions; manual appointment actions and periodic actions enter the same action pool, but retain different sources and statuses to avoid automatic reconstruction or duplicate reminders.

[0015] Furthermore, the method supports localized deployment for enterprises, with each enterprise deploying an independent instance. The CRM database, semantic processing model, attachments, original multimodal evidence, and audit logs are all stored locally on the enterprise's premises. Before message transmission, sensitive information that is not necessary for business purposes is anonymized on the client side. The chat software only handles message transmission and robot interaction and does not have the authority to directly write to the CRM. The semantic processing model can be replaced with a local or cloud-compatible service through a unified Provider interface, but regardless of the model's source, it is always subject to the rule that "the model only produces unconfirmed action plans, and the CRM Core makes the final decision."

[0016] According to another aspect of the present invention, a CRM information intelligent processing system based on natural language dialogue is provided. The system is implemented based on the above-mentioned CRM information intelligent processing method based on natural language dialogue, and specifically includes: a cross-platform message adaptation module, a multimodal semantic recognition module, a customer matching and deduplication module, a draft interaction confirmation module, a CRM business management module, a business analysis service module, and a data traceability module. The cross-platform message adaptation module connects to multiple mainstream chat software and receives various messages sent by employees. It performs message format standardization conversion, binds and maps chat accounts with CRM employee accounts, filters duplicate messages, and synchronously loads the corresponding employee's CRM data operation permissions. The multimodal semantic recognition module performs multimodal text parsing on the content of different message types, converting images, voice, and files into parsing evidence with original local references, time segments, page numbers, or image region locations, and performs comprehensive recognition based on the parsing evidence, current session context, employee identity, and CRM existing data. It extracts CRM business fields and identifies employee operation intentions. When the recognition results have missing fields, conflicting evidence, or confidence levels below the threshold, it generates natural language supplementary questions and sends them through the original session. The customer matching and deduplication module generates an unconfirmed action plan based on the extracted business fields. The action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependency and idempotent key. It also retrieves matching candidate customers in the CRM database and generates a structured CRM operation draft of the corresponding operation type after completing customer deduplication verification. The draft interaction confirmation module feeds back the operation draft to the original chat session, receives modification, confirmation, or cancellation instructions sent by employees via natural language, and updates the draft content. The confirmation instruction must simultaneously meet the following conditions: it comes from the same enabled robot connection, it comes from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and it points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies. After writing, a structured receipt based on the actual execution result is generated. For duplicate delivery of the original message, it is deduplicated by a combination of the platform identifier and the unique message identifier, and only existing drafts are reused. For duplicate sending of confirmation messages, it is deduplicated by a combination of the local instance, robot connection, and the unique confirmation message identifier, and only the first execution result is returned, without duplicate writing. The CRM business management module provides management functions for customer files, contacts, follow-up items and various business records. It supports customer search and filtering, batch transfer and archiving and import of external CRM data. It also includes a customer level-driven periodic follow-up engine, which maintains a unique pending periodic action for each customer who has enabled periodic follow-up. The countdown is reset only when the employee has clearly completed the follow-up and submitted new content. The business analysis service module aggregates CRM business data according to the company, department, and employee dimensions, generates periodic business analysis reports, and supports drill-down from statistical indicators to the corresponding business details and original messages. The data traceability module stores the original message, multimodal parsing results, semantic recognition process, customer matching details, draft version, confirmation record and final writing result, thus constructing a full-process data traceability link.

[0017] The beneficial effects of the CRM information intelligent processing system and method based on natural language dialogue of the present invention are as follows: By connecting to multiple mainstream instant messaging platforms and establishing a unified message standard and identity permission mapping, employees can use natural language to complete customer entry, follow-up records and information queries in commonly used chat software without having to frequently switch systems or manually fill out structured forms. At the same time, data submitted from different platforms can be automatically collected under the same employee account, effectively avoiding information dispersion and duplicate filing. In addition, through multi-round conversational incremental field completion and progressive intent recognition, it can better fit natural communication habits. When a field is missing, it can automatically guide the completion instead of directly reporting an error. Combined with the three-stage verification mechanism of draft preview, manual confirmation and formal writing, it can significantly reduce the direct impact of AI recognition bias on data. Furthermore, with the help of full-link traceability capabilities, it can support reverse tracing from the final business record to the original chat message, which is convenient for data verification and error correction. Meanwhile, operational data can be statistically analyzed in real time as needed and traced back to the original communication context, enabling managers to keep abreast of business dynamics. Local storage of core data and desensitization of sensitive information on the client side further strengthen data security and compliance capabilities. Attached Figure Description

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0019] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a schematic diagram of the system framework. Detailed Implementation

[0020] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0021] Example 1

[0022] According to one aspect of the invention, such as Figure 1 As shown, a CRM information intelligent processing method based on natural language dialogue is provided, including the following steps: Step 1: Message Access and Standardization By adapting to various mainstream chat software interfaces, it can receive various messages sent by employees, perform standardized message format conversion, bind and map chat accounts with CRM employee accounts, and filter duplicate messages.

[0023] First, corresponding adaptation service links are configured for mainstream chat software such as WeChat Work, DingTalk, and Lark. Long connection listening channels are established through the message push interfaces officially opened by each platform. When an employee sends text, image, voice or file messages to the system's dedicated account in the corresponding chat software, the corresponding platform pushes the message event to the system's adaptation service entry point to complete message reception and basic format verification, and discards invalid messages with abnormal formats. Then, the heterogeneous message data returned by different platforms is mapped to a unified standard message structure. The platform identifier, message unique number, session number, sender's external account, sending timestamp, message type and corresponding content and attachment metadata fields are extracted and organized in sequence. At the same time, the local storage reference and verification value of the original media file are recorded. Meanwhile, the pre-stored account binding relationship table is queried according to the sender's external account to match the corresponding CRM internal employee account, and the preset data access and operation permission scope of the account is loaded to limit the customer data boundary that can be accessed in subsequent business processing. Finally, the combination of the platform identifier and the unique message number is used as the deduplication verification key. The set of message identifiers that have been processed in the local cache is queried. If the verification key already exists, it is determined to be a duplicate push message, and the subsequent processing process is terminated directly. If the verification key does not exist, it is written to the cache and marked as pending processing. The standardized message data is then passed to the next processing stage.

[0024] Step 2: Parsing and Intent Recognition Multimodal text parsing is performed on different message types, converting images, audio, and files into parsing evidence with original local references, time segments, page numbers, or image region locations. Based on the parsing evidence, current session context, employee identity, and existing CRM data, comprehensive identification is performed to extract CRM business fields and identify employee operation intentions. When the identification results have missing fields, conflicting evidence, or confidence levels below the threshold, natural language supplementary questions are generated and sent through the original session to guide employees to supplement information.

[0025] The process begins by reading the message type identifier from standardized messages. Non-text messages are then categorized and converted into text: For image messages, a text recognition engine extracts text content region by region, recording the coordinates of each text fragment within the image. Simultaneously, a visual model identifies customer information, contacts, requirements, times, and business matters in business cards, product photos, quotations, and chat screenshots, labeling the identified names, contact information, company names, and their location. For voice messages, a speech-to-text engine converts audio content into time-stamped text, preserving the timestamp range and confidence level of each transcribed fragment. For file messages, the document content is parsed, extracting text fragments related to customer information and business communication, and recording the page number and paragraph position of each field. After conversion, historical message text from the same session is retrieved using the session identifier as an index. The text obtained from this conversion is then concatenated with the historical text in ascending order of sending timestamps to form a complete text string containing the full context for analysis. The CRM business field extraction and initial judgment of operation intent are based on the concatenated text string to be analyzed. Combining the current employee's identity and permission scope with the existing customer data in the CRM database, entity extraction and intent classification are performed through a domain-adapted semantic recognition model (such as BERT-base-Chinese). In the entity extraction stage, nine core business fields are identified in sequence: customer name, contact information, contact method, customer stage, demand description, budget range, purchase intention, follow-up content, and follow-up time. During the extraction process, historical field values ​​confirmed in the conversation are automatically reused, and only the fields added or adjusted in this round are updated. If there is a conflict in the field description, the latest message content shall prevail and the change record shall be retained. At the same time, each identified field is associated with its source evidence (specific text, image area, voice time segment or file paragraph). In the intent classification stage, the operation is first determined to belong to three categories: add, modify, and query. Then, it is further matched with specific operation types such as creating a customer, supplementing customer information, adding a contact, recording business events, arranging follow-up, and querying customer information. The confidence value of the corresponding intent is output simultaneously. Next, the extracted key fields such as customer name and contact information are initially correlated and verified with the existing CRM customer data within the employee's authority to verify the business rationality of the field content. At the same time, the completeness of the extracted core fields is calculated to determine whether there are any missing key fields and whether the confidence of the operation intention reaches the preset judgment threshold. If there are missing key fields or low confidence of the intention, or if there are conflicts between different sources of evidence (such as the date in the picture is inconsistent with the date in the voice), a corresponding natural language confirmation script is generated to clearly mark the information items to be supplemented or the conflict items. The employee is guided to gradually supplement or clarify through the original chat conversation. If the field completeness meets the standard and the confidence of the intention meets the threshold requirements, the current extraction results and intention judgment are confirmed to be valid. Finally, the confirmed set of valid business fields, their evidence location information, operation intent determination results, session context identifiers, and employee identity information are encapsulated to form a standardized semantic parsing result data package. The intermediate records of entity extraction and intent calculation details during this identification process are retained simultaneously, and the complete data package is passed to the next processing stage.

[0026] Step 3: Customer Matching and Draft Generation Based on the extracted business fields, an unconfirmed action plan is generated. This action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependencies, and idempotent keys. Matching candidate customers are retrieved from the CRM database. After deduplication verification, a structured CRM operation draft of the corresponding operation type is generated. The unconfirmed action plan is generated and output by the semantic processing model. The semantic processing model has no right to modify the draft status or call the CRM database write interface.

[0027] The process begins by extracting core identifier fields such as customer name, contact person name, and contact information from the semantic parsing results. Then, within the scope of CRM customer data accessible to the current employee, candidate customer recall is carried out. During recall, fields such as name keywords, mobile phone number, and landline number are used as search conditions to quickly filter out candidate customer entries with overlapping fields, thus narrowing down the scope of subsequent precise matching. Subsequently, a multi-dimensional matching degree calculation is performed on each candidate customer, calculating the text similarity of the customer name, the number of matching contact information, the degree of historical business association, and the degree of fit between the industry and region. Finally, the results are weighted and summarized according to preset weights to obtain the final matching score. At the same time, the detailed calculation of each dimension is recorded, and the matching process data is completely retained. The tiered processing of customer matching results and the confirmation of operation targets are performed based on the preset range of the final matching score: when the matching score reaches the high confidence range, it is directly determined to be the same customer, and the existing customer file is identified as the associated object for this operation. When the matching score is in the medium confidence range, the key information such as the name of each candidate customer, core contact person, and most recent follow-up time is compiled to form a concise candidate list, which is ready to be pushed to the original chat session for employees to select and confirm. When the matching score is in the low confidence range, it is determined that the current customer information is not in the existing data, and the operation of creating a new customer file is executed. All matching judgment results are simultaneously marked with the corresponding confidence level and judgment basis, and flow down with the business data. When generating an action plan, the identified operational intent, confirmed customer association results, and all extracted business fields are mapped and populated into the structured draft template of the corresponding operation type. At the same time, action batches are formed according to the logical relationship between operations (e.g., the dependency between "creating a new customer" and "adding a contact under this customer"). Each action is assigned a draft version identifier and an idempotent key, and the operator information is recorded. For the "creating a customer" operation, a complete customer profile draft is generated. For operations such as supplementing information, adding contacts, recording business events, and arranging follow-ups, draft content of corresponding business records is generated. After the draft is populated, the format validity of all fields is automatically checked. The date format, contact information format, values ​​of enumerated fields, and ranges of numeric fields are checked to see if they conform to CRM data specifications. Fields that fail the check are automatically marked with an error message. After the check is completed, a temporary operation draft with a version identifier is generated, associated with the corresponding session identifier and employee identity information, and temporarily stored in the temporary storage area for subsequent confirmation. At the same time, the complete process record of draft generation is retained.

[0028] Step 4: Draft Confirmation and Data Writing The operation draft is fed back to the original chat session, receiving modification, confirmation, or cancellation instructions sent by employees via natural language. The draft content is updated, and the confirmation instruction must simultaneously meet the following conditions: from the same enabled robot connection, from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and it points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies. Within the same business transaction, the confirmation status change, formal data writing, audit record generation, and structured receipt generation are executed atomically. If any step fails, the entire process is rolled back, and no partially written formal objects are generated. After writing, a structured receipt based on the actual execution result is generated, including the specific objects created or updated and the follow-up time. The structured receipt is generated by the CRM Core based on the actual writing result. The semantic processing model has no right to generate or modify the receipt content, nor does it have the right to generate a conclusion indicating successful writing.

[0029] First, the generated structured operation draft with version identifier is converted into an easy-to-read display format. The corresponding customer information, contact information, follow-up content or business event details are sorted according to the operation type and presented in a clear itemized form. Then, based on the session identifier and the original platform identifier, the message reply interface of the corresponding chat software is called to push the draft preview content to the original chat session in which the employee initiated the request for verification. After the draft is pushed, it is marked as pending confirmation and continues to be kept in the temporary storage area. Before receiving the employee's confirmation instruction, the formal write operation to the CRM database is never triggered. After receiving the instruction message from the employee in the original session, the system first performs semantic parsing on the instruction content to determine whether the employee's operation intention is to modify, confirm, or cancel. If it is determined to be a modification instruction, the system further parses the target fields and corresponding modification values ​​involved in the instruction, locates the corresponding fields in the current draft, completes the content update, updates the draft version number synchronously, and re-executes the field format validity check. After the check passes, the draft preview content is generated again and pushed back to the session for the employee to check again. It supports multiple rounds of iterative modification until the employee confirms or cancels. If it is determined to be a cancellation instruction, the system directly clears the corresponding draft data in the temporary storage, terminates the current operation process, and feeds back the cancellation result to the employee. Once the parsing determines that a clear confirmation instruction has been received, the system does not immediately execute the write operation, but instead performs four binding checks: ① The confirmation message must come from the same enabled robot connection (i.e., a connection that is active in the channel adaptation layer). ② The confirmation message must come from the same session that generated the draft (i.e., the session identifiers must be the same); ③ The CRM employee account obtained by mapping the sender's external account through the binding relationship must be consistent with the operator of the draft record or an authorized manager; ④ The draft version number indicated in the confirmation message must be completely consistent with the currently valid version number in temporary storage. If any condition is not met, the system will refuse to execute and return the specific reason for failure. After successful verification, the CRM Core (running independently of the semantic processing model and performing verification and formal writing as a controlled service) retrieves the latest version of the draft data, completes data mapping and transformation according to the table structure specifications of the CRM database, and performs formal data writing operations in the same business transaction, generating corresponding customer profiles, contact records, follow-up records, or business event records. At the same time, it verifies field rules, object dependencies, and transaction integrity. If any step fails, the entire process is rolled back, and no incomplete formal objects are generated. After the write is completed, the successful operation result, corresponding customer name, operation type and key information are organized into a structured receipt based on the actual execution result (clearly listing which specific records were created or updated, which customer is associated with, and when follow-up is required). This receipt is then pushed back to the original chat session through the original chat software interface to inform the employee. At the same time, the employee confirmation instruction record, the final draft version and the official write result are retained and linked to the corresponding traceability link to complete the closed loop of this conversational CRM operation.

[0030] Step 5: Statistics and Source Tracing CRM business data is aggregated and statistically analyzed by company, department, and employee to generate periodic business analysis reports. At the same time, the data of the entire process operation nodes is stored to build a data traceability link. For duplicate delivery of original messages, the platform identifier and the unique message identifier are combined to remove duplicates, and only existing drafts are reused. For duplicate sending of confirmation messages, the local instance, robot connection, and the unique confirmation message identifier are combined to remove duplicates, and only the first execution result is returned, without duplicate writing.

[0031] The process begins by extracting full business data, including customer profiles, follow-up records, and business events, from the CRM business database within a specified time frame. This data is then aggregated and calculated according to three organizational dimensions: company as a whole, subordinate departments, and individual employees. Key operational indicators such as the number of new customers, customer stage distribution, number and rate of completed follow-up plans, number of overdue and unprocessed matters, and business dynamics of high-level customers are then statistically analyzed for each dimension. After the statistics are completed, a mapping relationship between the indicators and the corresponding business data is established. This allows for step-by-step queries from any statistical indicator, first locating the corresponding customer list from the indicator, then entering the details of the corresponding business record from the customer profile, and finally jumping to view the original chat messages and attachments that generated the business data. This enables full-link traceability of operational data from macro to micro levels. The establishment of a full-link data traceability system and the retention of data at all nodes involves assigning a unique traceability identifier to each CRM operation initiated through chat software. This identifier is used as the primary key to connect and store all node data throughout the entire operation process. The system sequentially retains the original message text and attachments, intermediate results and evidence location information (image areas, voice time segments, file page numbers) from multimodal content parsing, entity extraction records and intent determination details during semantic recognition, candidate lists and score calculation processes in the customer matching stage, content and change records of each version of the operation draft, confirmation or modification instruction records sent by employees, and the execution results finally written to the database. All traceability data is stored by appending, supporting only addition and query operations, but not modification or deletion, ensuring that every piece of CRM business data can be traced back to the original communication source, meeting the needs of business verification and compliance audit. In addition, the system has a built-in customer level-driven periodic follow-up mechanism: it maintains a unique pending periodic action for each customer who has enabled periodic follow-up, and the customer level corresponds to a configurable follow-up frequency (e.g., once a week for level A, once every two weeks for level B). Only when the employee clearly completes the follow-up and submits new communication content (chat logs, voice, pictures or manual descriptions) will the system recalculate the countdown for the next round based on the actual completion time and update the status of the periodic action. Manually scheduled actions (such as "call back next Wednesday") and recurring actions are placed in the same action pool, but retain different source identifiers and statuses. Scheduled tasks are only responsible for summarizing actions due on the day and sending reminders. The same to-do list must not be created repeatedly, nor can existing countdown actions be automatically deleted or replaced due to level adjustments. Finally, according to the preset daily, weekly, and monthly reporting cycles, the system automatically aggregates the corresponding operational statistics for each cycle, compiling them into periodic operational reports that include an operational overview, key reminders, and trends in core indicators. These reports are proactively pushed to the corresponding manager accounts via chat software. Simultaneously, a traceability query capability is provided, allowing managers and auditors to retrieve the corresponding traceability identifiers through customer records, operation times, or employee accounts, and to verify the complete operational process node data. Combined with a localized deployment architecture: each enterprise deploys an independent instance; the CRM database, semantic processing model, attachments, original multimodal evidence, and audit logs are all stored on the enterprise's local server; the chat software only handles message transmission and robot interaction and does not have direct write permissions to the CRM; the semantic processing model can be replaced with a local llama-server or a cloud-compatible service through a unified Provider interface, but regardless of the model's source, it is always constrained by the rule that "the model only produces unconfirmed action plans, and the CRM Core makes the final decision." Simultaneously, sensitive information is de-identified locally on the client side before message transmission (e.g., identifying and blocking non-business-essential sensitive fields such as ID card numbers and bank card numbers on the client side before sending messages to ensure that sensitive data does not enter the network transmission link). This ensures that data is traceable and analyzable while comprehensively protecting the security and compliance of enterprise customer data.

[0032] Example 2

[0033] This invention also provides a CRM information intelligent processing system based on natural language dialogue. This system corresponds to the overall framework provided in the foregoing embodiments, and is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 2 As shown, the system mainly includes: a cross-platform message adaptation module, a multimodal semantic recognition module, a customer matching and deduplication module, a draft interaction confirmation module, a CRM business management module, a business analysis service module, and a data traceability module; The cross-platform message adaptation module integrates with multiple mainstream chat software and receives various messages sent by employees. It performs standardized message format conversion, binds and maps chat accounts with CRM employee accounts, filters duplicate messages, and synchronously loads the corresponding employee's CRM data operation permissions. The multimodal semantic recognition module performs multimodal text parsing on the content of different message types, converting images, voice, and files into parsing evidence with original local references, time segments, page numbers, or image region locations, respectively. Based on the parsing evidence, the current session context, employee identity, and existing CRM data, it performs comprehensive recognition, extracts CRM business fields, and identifies employee operation intentions. When the recognition results have missing fields, conflicting evidence, or confidence levels below the threshold, it generates natural language supplementary questions and sends them through the original session. The customer matching and deduplication module generates an unconfirmed action plan based on the extracted business fields. The action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependency and idempotent key. It also retrieves matching candidate customers from the CRM database and generates a structured CRM operation draft for the corresponding operation type after completing the customer deduplication verification. The draft interaction confirmation module feeds back the operation draft to the original chat session, receives modification, confirmation, or cancellation instructions sent by employees via natural language, and updates the draft content. The confirmation instruction must simultaneously meet the following conditions: it comes from the same enabled robot connection, it comes from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and it points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies. After writing, a structured receipt based on the actual execution result is generated. For duplicate delivery of the original message, it is deduplicated by a combination of platform identifier and message unique identifier, and only existing drafts are reused. For duplicate sending of confirmation messages, it is deduplicated by a combination of local instance, robot connection, and confirmation message unique identifier, and only the first execution result is returned, without duplicate writing. The CRM business management module provides management functions for customer files, contacts, follow-up items and various business records. It supports customer search and filtering, batch transfer and archiving and import of external CRM data. It also includes a customer level-driven periodic follow-up engine, which maintains a unique pending periodic action for each customer who has enabled periodic follow-up. The countdown is reset only when the employee has clearly completed the follow-up and submitted new content. The business analysis service module aggregates CRM business data according to company, department, and employee dimensions, generates periodic business analysis reports, and supports drill-down from statistical indicators to corresponding business details and original messages. The data traceability module stores the original message, multimodal parsing results, semantic recognition process, customer matching details, draft version, confirmation record and final write result, and builds a full-process data traceability link.

[0034] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also within the protection scope of the present invention.

Claims

1. A CRM information intelligent processing method based on natural language dialogue, characterized in that, The method includes: S1. Connect to various mainstream chat software through the adapter interface, receive various messages sent by employees, perform standardized message format conversion, bind and map chat accounts with CRM employee accounts, and filter duplicate messages; S2. Perform multimodal text parsing on the content of different message types, converting images, voice, and files into parsing evidence with original local references, time segments, page numbers, or image area locations, and perform comprehensive identification based on the parsing evidence, current session context, employee identity, and CRM existing data. Extract CRM business fields and identify employee operation intentions. When the identification results have missing fields, conflicting evidence, or confidence levels below the threshold, generate natural language supplementary questions and send them through the original session to guide employees to supplement information. S3. Generate an unconfirmed action plan based on the extracted business fields. The action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependency and idempotent key. Retrieve matching candidate customers in the CRM database. After completing customer deduplication verification, generate a structured CRM operation draft of the corresponding operation type. The unconfirmed action plan is generated and output by the semantic processing model. The semantic processing model has no right to modify the draft status or call the CRM database write interface. S4. Feedback the operation draft to the original chat session, receive modification, confirmation, or cancellation instructions sent by employees via natural language, update the draft content, and the confirmation instruction must simultaneously meet the following conditions: from the same enabled robot connection, from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies, and atomically executes the confirmation status change, formal data writing, audit record generation, and structured receipt generation in the same business transaction. If any step fails, the entire process is rolled back, and no partially written formal objects are generated. After writing, a structured receipt based on the actual execution result is generated, which includes the specific objects created or updated and the follow-up time. The structured receipt is generated by the CRM Core based on the actual writing result. The semantic processing model has no right to generate or modify the receipt content, nor does it have the right to generate a conclusion indicating that the writing was successful. S5 aggregates CRM business data by company, department, and employee, generates periodic business analysis reports, and stores data of all operation nodes to build a data traceability chain. For duplicate delivery of original messages, it uses a combination of platform identifier and message unique identifier to deduplicate, and only reuses existing drafts. For duplicate sending of confirmation messages, it uses a combination of local instance, robot connection, and confirmation message unique identifier to deduplicate, and only returns the first execution result, without writing it repeatedly.

2. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: In step S1, the standardized and transformed message includes platform identifier, message unique identifier, session identifier, sender account, sending time, message type, and content attachment information; after account binding and mapping, the corresponding employee's CRM data operation permissions are loaded synchronously to limit the range of customer data that employees can access; the combination of platform identifier and message unique identifier is used as the deduplication key to intercept duplicate messages.

3. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: In step S2, messages of image, voice, and file types are converted into text content through text recognition, speech-to-text, and document parsing, respectively, while retaining the original local references, timestamps or regional coordinates of the recognized segments, and page number information. These are then concatenated with the plain text content in the conversation in chronological order. The semantically extracted business fields include customer name, contact information, contact method, customer stage, needs, budget, purchasing intention, follow-up content, and follow-up time. The recognized operation intentions include creating a customer, supplementing customer information, adding a contact, recording business events, arranging follow-up, and querying customer information.

4. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: In step S3, customer matching uses multi-feature weighted calculation to calculate the matching score, and performs automatic association, in-session manual confirmation, or new customer processing according to the score classification. The matching process does not rely on word frequency-inverse document frequency weight or n-gram weight models trained on historical unstructured data, does not use knowledge graph triple inference as the matching path, and does not use automatic contact creation by phone number as the main link. The structured operation draft contains customer profiles to be written, contact information, follow-up records, and business event details, and completes field format validity checks. The action dependencies in the unconfirmed action plan support the arrangement of multiple actions into the same transaction batch, such as creating a customer first, and then creating a contact and follow-up action under that customer.

5. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: In step S4, the operation draft is pushed to the original chat session in the form of structured cards or natural language summaries; when the employee's modification instruction is received, the corresponding fields and values ​​are parsed and the draft is updated and then pushed back for preview; if no confirmation instruction is received, the draft is only temporarily cached and is not written to the CRM database. Once the write operation is complete, the execution result will be restored to the original chat session. The semantic processing model does not hold CRM database write credentials; the CRM Core runs independently of the semantic processing model, and receives action plans and performs verification and submission through internal service interfaces as a controlled service.

6. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: In step S5, the statistical business data includes the number of new customers, changes in customer stages, follow-up completion status, overdue matters, and dynamics of high-level customers; it supports drill-down from statistical indicators to the corresponding customers, business records, and original chat messages; the node data stored throughout the process includes original messages, multimodal parsing results and evidence location, semantic recognition process, customer matching details, draft versions, confirmation records, and final write results.

7. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: The method also supports a customer-level driven periodic follow-up mechanism: a unique pending periodic action is maintained for each customer who has enabled periodic follow-up, and the customer level corresponds to a configurable follow-up frequency; the system recalculates the next countdown based on the actual completion time only after the employee has clearly completed the follow-up and submitted new communication content or manual instructions; manual appointment actions and periodic actions enter the same action pool, but retain different sources and statuses to avoid automatic reconstruction or duplicate reminders.

8. The intelligent CRM information processing method based on natural language dialogue according to claim 1, characterized in that: The method supports localized deployment for enterprises, with each deployment using an independent instance. The CRM database, semantic processing model, attachments, original multimodal evidence, and audit logs are all stored locally on the enterprise's premises. Before message transmission, sensitive information that is not necessary for business purposes is anonymized on the client side. The chat software only handles message transmission and robot interaction and does not have the permission to directly write to the CRM. The semantic processing model can be replaced with a local or cloud-compatible service through a unified Provider interface, but regardless of the model's source, it is always subject to the rule that "the model only produces unconfirmed action plans, and the CRM Core makes the final decision".

9. A CRM information intelligent processing system based on natural language dialogue, characterized in that, The system is implemented based on a natural language dialogue-based intelligent CRM information processing method as described in any one of claims 1-8, specifically including: a cross-platform message adaptation module, a multimodal semantic recognition module, a customer matching and deduplication module, a draft interaction confirmation module, a CRM business management module, a business analysis service module, and a data traceability module; The cross-platform message adaptation module connects to multiple mainstream chat software and receives various messages sent by employees. It performs message format standardization conversion, binds and maps chat accounts with CRM employee accounts, filters duplicate messages, and synchronously loads the corresponding employee's CRM data operation permissions. The multimodal semantic recognition module performs multimodal text parsing on the content of different message types, converting images, voice, and files into parsing evidence with original local references, time segments, page numbers, or image region locations, and performs comprehensive recognition based on the parsing evidence, current session context, employee identity, and CRM existing data. It extracts CRM business fields and identifies employee operation intentions. When the recognition results have missing fields, conflicting evidence, or confidence levels below the threshold, it generates natural language supplementary questions and sends them through the original session. The customer matching and deduplication module generates an unconfirmed action plan based on the extracted business fields. The action plan contains at least one CRM action. Each action includes object type, operation type, fields to be written and their source evidence, candidate customer list, draft version identifier, operator information, action dependency and idempotent key. It also retrieves matching candidate customers in the CRM database and generates a structured CRM operation draft of the corresponding operation type after completing customer deduplication verification. The draft interaction confirmation module feeds back the operation draft to the original chat session, receives modification, confirmation, or cancellation instructions sent by employees via natural language, and updates the draft content. The confirmation instruction must simultaneously meet the following conditions: it comes from the same enabled robot connection, it comes from the same session that generated the draft, the sender is mapped to the draft owner or authorized personnel, and it points to the current valid draft version. Only when all conditions are met will the draft data be written to the CRM database. Before writing, the CRM Core independently reviews the source legality, operation permissions, field rules, and object dependencies. After writing, a structured receipt based on the actual execution result is generated. For duplicate delivery of the original message, it is deduplicated by a combination of the platform identifier and the unique message identifier, and only existing drafts are reused. For duplicate sending of confirmation messages, it is deduplicated by a combination of the local instance, robot connection, and the unique confirmation message identifier, and only the first execution result is returned, without duplicate writing. The CRM business management module provides management functions for customer files, contacts, follow-up items and various business records. It supports customer search and filtering, batch transfer and archiving and import of external CRM data. It also includes a customer level-driven periodic follow-up engine, which maintains a unique pending periodic action for each customer who has enabled periodic follow-up. The countdown is reset only when the employee has clearly completed the follow-up and submitted new content. The business analysis service module aggregates CRM business data according to the company, department, and employee dimensions, generates periodic business analysis reports, and supports drill-down from statistical indicators to the corresponding business details and original messages. The data traceability module stores the original message, multimodal parsing results, semantic recognition process, customer matching details, draft version, confirmation record and final writing result, thus constructing a full-process data traceability link.

Images